fix(kimi-coding): never send thinking{type:disabled} on the api.kimi.com/coding endpoint

[Yackadaisical]

Problem

The Anthropic-Messages-shaped Kimi coding endpoint (api.kimi.com/coding) rejects thinking: {"type": "disabled"} with:

HTTP 400: invalid thinking: only type=enabled is allowed for this model

KimiProfile.build_api_kwargs_extras emits {"type": "disabled"} whenever a turn requests reasoning-off (reasoning_config={"enabled": False}). Any cron/agent run that disables reasoning mid-session therefore 400s intermittently — the default reasoning_effort path sends a top-level reasoning_effort instead and is unaffected, so it only fails on the disable-reasoning turns.

Observed in production on a daily cron using provider: kimi-coding + base_url: https://api.kimi.com/coding: most days errored, days that never disabled reasoning succeeded.

Fix

Detect the /coding endpoint from the resolved base_url and degrade a disable request to the only value that endpoint accepts, {"type": "enabled"} (thinking cannot be turned off there anyway). The OpenAI-compat /v1 endpoint keeps emitting disabled unchanged, so no behavior changes for Moonshot /v1 users.

Tests

Adds test_coding_endpoint_never_disables_thinking; all existing kimi-profile assertions still pass.

[teknium1]

Thanks for the focused regression report and test.

Automated hermes-sweeper review found that this behavior is already implemented on current main:

• hermes_cli/runtime_provider.py:137-138 routes api.kimi.com/coding to anthropic_messages, rather than the Chat Completions profile path changed here.
• agent/anthropic_adapter.py:2641-2643 suppresses the Anthropic thinking parameter for Kimi endpoints, including an explicit disabled-reasoning request.
• tests/agent/test_kimi_coding_anthropic_thinking.py:56-67 already covers reasoning_config={"enabled": false} for https://api.kimi.com/coding and asserts that thinking is absent.
• The behavior was shipped by 410f33a728bbe97b4853e6f411c8a1d835d7de9c (fix(kimi): don't send Anthropic thinking to api.kimi.com/coding, fix(kimi): don't send Anthropic thinking to api.kimi.com/coding #13826).

Closing as already implemented on main.